oRGB: A Practical Opponent Color Space for Computer Graphics

Transcription

oRGB: A Practical Opponent Color Space for Computer Graphics
oRGB: A Practical Opponent Color Space for Computer Graphics
Margarita Bratkova∗
Solomon Boulos†
University of Utah
Peter Shirley‡
Figure 1: Visualization of the oRGB decomposition of the image on the left into luma, yellow-blue, and red-green channels.
Abstract
We present a new color model, oRGB, that is based on opponent
color theory. Like HSV, it is designed specifically for computer
graphics. However, it is also designed to work well for computational applications such as color transfer, where HSV falters. Despite being geared towards computation, oRGB’s natural axes facilitate HSV-style color selection and manipulation. oRGB also allows for new applications such as a quantitative cool-to-warm metric, intuitive color manipulations and variations, and simple gamut
mapping. This new color model strikes a balance between simplicity and the computational qualities of color spaces such as CIE
L*a*b*.
1
Introduction
Digital images are stored, manipulated, and transmitted in a variety of color spaces. Many different color spaces have been proposed, with RGB, CMYK, HSV, and CIE L*a*b* being especially
popular (see [Hunt 2004; Fairchild 2005]). The RGB space is
an additive color model and is probably the most commonly used
among computer graphics practitioners, but it is not very intuitive.
CMYK is a subtractive color model and is most useful in color printing. A bijective transform of RGB space, HSV was introduced by
Smith [Smith 1978] as a more convenient and meaningful color
space, designed specifically for computer graphics. At about the
same time, the need for a device independent, perceptually linear
color space led to the development of CIE L*a*b*.
The recent interest in automatic image manipulation techniques
such as color harmonization, color transfer, colorization, as well
as the proliferation of work in computational photography, suggest
a need for an intuitive, yet simple and practical computational computer graphics color space. While HSV has proven extremely useful
in practice, there are two characteristics it lacks from our standpoint. First, its primary colors (red, green, and blue) are based on
the mechanics of additive color mixing rather than the more natural
psychological primary colors. Second, HSV is well-suited for color
selection, but is not a good space for computational applications.
In this paper we introduce a new color space, the oRGB model.
Like HSV, it is a an invertible transform from RGB. The primaries
∗ [email protected][email protected][email protected]
of this model are based on the three fundamental psychological opponent axes (white-black, red-green, and yellow-blue). The model
has three channels - one luma (see Section 2.5 for a definition) and
two chroma, as shown in Figure 1 . The non-linearity of these channels relative to intensity uses the non-linear gamma encoding of the
native RGB space. Thus it has some of the advantages of non-linear
perceptual spaces such as CIE L*a*b*, while maintaining a simple gamut and computationally simple transform to and from RGB.
Because the chroma axes of oRGB go from warm to cool (red to
green and yellow to blue), oRGB has a quantitative concept of color
warmth as needed by some artistic applications. This new space is
reasonably simple, has intuitive axes, and is well-suited for computational applications.
In the rest of this work, we first review opponent process color theory and some of the color models used in computer graphics (Section 2). We then develop the details of the oRGB model (Section 3),
and show the application of this color space for color adjustment,
NPR lighting computation, color transfer, and gamut mapping (Section 4).
2
Background
We now review the high-level differences between various color ordering systems, paying particular attention to opponent color systems. We also establish several terms and symbolic conventions as
the study of color spans several disciplines and there are no universally consistent uses.
2.1
Hue Spacing Systems
Most color spaces have one light-dark axis and two chromatic axes.
The boundary of the realizable colors in the chromatic dimension is
often displayed as a circle or “wheel”. The ordering of hues is the
same in most color spaces, but the spacing between the hues differs.
We summarize these to clarify the relationship between opponent
color systems and more traditional hue ordering systems.
Four of the most popular types of color circles are shown in Figure 2. The two circles to the left are based on three primaries and
how they mix. The uniform color circle, of which the one that visualizes the Munsell Color Space is perhaps the best known, attempts to arrange the hues so they are “evenly spaced”. Such uniform spaces often lack a concept of “primaries” as uniformity precludes even spacing between named colors such as red and blue.
To the right is a Hering-style opponent circle with four primaries
on two natural axes. It emphasizes his concept of red-green and
yellow-blue axes, while maintaining the order of colors as in the
other color circles.
2.2
all of these models have more of a magenta-cyan axis than a redgreen one. An opponent space proposed for computer graphics by
Naiman [Naiman 1985] is similar to YCC (thought it has a slightly
different matrix, and operates on linear RGB), but it also has an axis
shifted toward magenta-cyan.
Hering’s Opponent Process Color Theory
The opponent process theory of color was first advocated by Ewald
Hering in the 1870s. It took a sharp departure from the prevalent
theory of the time, the trichromatic theory of Young-Helmholtz, by
proposing four hue primaries: red, green, yellow, and blue, instead
of the traditionally accepted three: red, green, and blue.
While three primaries are needed to produce all possible colors,
Hering noted that our subjective experience of color suggests that
perceptually there is an encoding for the fourth. For example, while
purple seems to be a mixture of red and blue, yellow does not seem
to be composed of any other primaries. Instead, it seems to be a
“pure” color, a primary itself. In addition, Hering observed that we
don’t seem to have a phenomenological experience for a color that
is both yellow and blue, nor red and green. However, we have one
for a color that is both yellow and red (orange).
These observations led Hering to hypothesize that human color perception uses two chromatic axes: one that ranges from red to green,
and another from yellow to blue. If these axes are independent, then
people will be able to see a red and yellow color simultaneously (orange), but not a red and green color, nor a blue and yellow color, as
they share the same “channel” (see Figure 3).
reds
blues
yellows
greens
Figure 4: Visualization of the chroma channels for (left to right):
YCC, l αβ , and CIE L*a*b color spaces. Compared to our new
oRGB space (Figure 1), these spaces do not have a true red-green
channel.
The most well-known opponent space with a true red-green axis
is the NCS. Unfortunately, there are no accepted color formulas to
allow this space to be used for computation [Kuehni 2004], and is
therefore not well suited for computer graphics applications.
We are aware of only three quantitative Hering-style opponent models. The first is that of Schwarz et al. [Schwarz et al. 1987] who
embed their space in XYZ tristimulus space with non-orthogonal
but true opponent axes. The second, introduced by Ware and
Cowen [Ware and Cowan 1990], places the opponent primaries at
corners of the model. Unfortunately, the axes are not opponent. The
third is that of Shirriff [Shirriff 1993] who developed an orthogonal red-green and yellow-blue space, using HSV as an intermediary
space. Each of these spaces were designed for color selection, are
defined in terms of linear RGB, and are not well suited for computational color applications. We will explore this further in Section 4.
2.4
Figure 3: The color circle divided into its opponent sides. The top
part of the circle is composed of colors that have a red tone in them,
the bottom - green, the right - yellow, and the left - blue.
Hering’s ideas were controversial at the time. However, there has
been a wealth of psychological and physiological evidence since
then (Hurvich [Hurvich and Jameson 1957] brought the revival of
the theory) that suggests that the organization of human color perception is at least partially based on opponent axes [Conway 2002;
Valberg 2005]. Color scientists have moved more and more towards
including some form opponency into many of their color spaces,
with CIE L*a*b* and the Natural Color System (NCS) being notable examples.
2.3
Opponent Space Models
A variety of spaces are based on opponent axes, with the television
encoding YIQ being perhaps the best known. However, we make
the distinction that most of these spaces are not Hering-style opponent spaces in that their principal axes do not include a red-green
dimension (while some of them do include a blue-yellow). This
is illustrated in Figure 4 for YCC (a modern digital descendant of
YIQ), l αβ (a color space based on natural image statistics), and CIE
L*a*b* (an approximately perceptually uniform space). Note that
Limitations of Color Models
One approach to generating a color model is to proceed directly
from existing scientific knowledge. While one might expect color
science to be both a mature and static field, this is far from the case.
The details of the physiological origin of color perception remain
enigmatic despite the increasing understanding of opponent mechanisms [Conway 2002; Valberg 2005]. The psychophysical approach that is used in the construction of most models also has limitations. There is unlikely to be a “right” choice, as different color
spaces have different goals and are all simplifications of the real experience [Kuehni 2003]. For this reason, existing color spaces are
based at least partially on heuristics and differ largely in their goals
and how they manage trade-offs.
Opponent color theory presents many unknowns and asymmetries.
For example, it is believed that the opponent signals are not simple
linear combinations of cone signals as is often diagrammed in textbooks [Chichilisky and Wandell 1999; Valberg 2005]. In addition,
the axes do not all combine in symmetrical ways. The combinations
of some axes create “important” colors with their own names, while
others do not [Hardin 1998]. For example, red and yellow make orange, while green and yellow make chartreuse (which most English
speakers associate with a pink or purplish color). Fundamental colors such as blue still look “blue” when darkened, while some may
change color category such as an orange turning to “brown” when
darkened. In addition, the chromatic axes have a natural symmetry
subtractive
additive
uniform
opponent
Figure 2: Visualization of four common types of color circles. The ordering of hues is preserved in each, while the position of the hues on
the circle varies. The subtractive circle is that used by painters while a cyan-magenta-yellow circle (not shown) is used by most ink printing
processes.
around grey while the achromatic black-white axis is believed to be
(and represented in most color models as) just positive [Volbrecht
and Kliegl 1998; Kuehni 2003].
These uncertainties and trade-offs suggest that there is no “true”
color space. It is important that one bases the choice of color space
on its performance in the applications of interest. Here we propose
a color space that is suitable for computational applications and is
simple, intuitive, and practical.
2.5
Terminology and Notation
Most computer graphics images are stored in some form of an RGB
color space. While there are many different RGB spaces, we assume that for the most part, our images are encoded in sRGB, an
increasingly popular standard. An often confusing notational issue
is distinguishing between linear and non-linear intensity red, green,
and blue components. A linear red channel stored with range [0, 1]
for example, would represent half the maximum physical intensity
with 0.5. A non-linear one, however, might do so at 0.75, depending on the coefficients used.
We adopt the often used convention that a prime (′ ) is used to denote
a quantity that is non-linear in intensity and ready for display. So
the non-linear R′ G′ B′ stored in most files contains (R′ , G′ , B′ ), while
a linear intensity RGB contains (R, G, B). For sRGB the relationship
between components is well approximated by R′ ≈ R1/2.2 . When
converting to linear photometric quantities such as (X,Y, Z), the
linear (R, G, B) should be used.
Linear color spaces usually use luminance, a linear photometric
standard for achromatic light response, and often denoted by the
symbols L and Y . The value for luminance can be computed as a
linear transform of a color value stored in linear color space. For
example, for linearized sRGB values, the luminance is:
L = 0.213R + 0.715G + 0.072B
(1)
Computing luminance as a linear combination of (R′ , G′ , B′ ) is not
possible. Instead, a non-linear analog of luminance called luma
is often used. There is no single standard for luma computation
(see Poyton [Poynton 2004] for an excellent detailed discussion of
different luma computation standards). Sometimes the symbols L
or Y are also used for luma, but we use L′ to distinguish it from
luminance. Typically, luma is a linear combination of (R′ , G′ , B′ ).
For example the YCC standard uses:
L′ = 0.299R′ + 0.587G′ + 0.114B′
(2)
The non-luma axes typically encode chromatic information, and are
also linear combinations of (R′ , G′ , B′ ). The color quantities are
usually called chroma.
The gamut of an RGB monitor is the set of colors it can physically
reproduce. In RGB space this gamut is the unit cube [0, 1]3 . When a
color space has a well-defined mapping from RGB, the RGB gamut
within that space is the shape created by sending the entire RGB
cube through that mapping. For some color spaces the boundaries
of that RGB gamut are algebraically simple, e.g. the RGB gamut of
the YIQ space is a parallelepiped. For others such as CIE L*a*b*
the surfaces on the RGB gamut boundary are algebraically complicated. When color modifications take colors outside the RGB
gamut, some method must be used to map these colors back into
gamut before they can be stored in typical RGB files. Such a method
is usually called gamut mapping (for an overview see [Morovic
and Luo 2001]) and component by component truncation is often
used [Stone et al. 1988].
3
The oRGB Color Space
As our interest is an opponent color space that is ideal for RGB
computation, we begin with Alvy Ray Smith’s original observation
in his classic HSV paper:
Full-blown color theory is a quite complex subject, involving physics, psychology, and physiology, but restriction to the RGB monitor gamut simplifies matters
substantially [Smith 1978].
Smith went on to design HSV as a simple bijective transformation
of RGB that is useful for color picking, as it has natural axes of
lightness, hue, and saturation. We would like an analogously simple space that retains these intuitive notions, but also uses the natural axes of light-dark, red-green, and yellow-blue as suggested by
perceptual theories. Unlike HSV, however, we would like to be able
to perform well in computational applications.
Inspired by Poyton’s framework [Poynton 2004] for how different
color models deal with CIE tristimulus values, we note how two
popular color spaces, CIE L*a*b* and l αβ operate. Both of these
spaces have been used effectively in computational settings, and unlike other opponent-like color spaces, store quantities that are nonlinear in intensity. However, both of these spaces have complicated
pipelines as shown in Figure 5. Each of them first “linearizes” the
(R′ , G′ , B′ ) to (R, G, B), does a linear transform to an intermediate
linear intensity space (either XYZ or LMS), applies a componentby-component non-linear map (either logarithmic or polynomial),
and then does a linear transform to get into an opponent-like space.
Our goal is to design a space that is useful for graphics applications. Having a non-linear luma channel is more perceptually uniform than luminance, and we believe it is simpler, improves computational speed, and also works better for practical applications (see
L*
a*
# # #
# # #
# # #
b*
linear
transform
l
α
# # #
# # #
# # #
β
linear
transform
polynomial
polynomial
polynomial
logarithm
logarithm
logarithm
X
# # #
# # #
# # #
Y
Z
linear
transform
L
G
B
R
# # #
# # #
# # #
M
S
R
linear
transform
G
B
polynomial
R'
polynomial
G'
polynomial
B'
polynomial
R'
polynomial
G'
polynomial
B'
Figure 5: The pipelines of R′ G′ B′ to CIE L*a*b* (top) and R′ G′ B′
to l αβ (bottom).
Figure 19 in Section 4.3). Therefore, instead of going through a
device-independent space and reintroducing non-linearity, we simply apply a linear transform directly to the non-linear R′ G′ B′ values
to get to an intermediate linear space we denote as L′C′C′ . Like
the CIE L*a*b* and l αβ chroma axes, L′C′C′ does not have true
Hering-style opponent axes. Therefore, we introduce a final rotation around the L′ axis that achieves a true Hering-opponent model
- oRGB (see Figure 6).
L'
L'
C'yb
C'rg
nonuniform
rotation
around
L' axis
Figure 6: The pipeline of
3.1
C'1
C'2
R′ G′ B′
G'
linear
transform
B'
to oRGB
′ ,C′ ).
(L′ ,Cyb
rg
oRGB Step 1: Linear transform of R′ G′ B′ to L′C′C′
There is an existing family of color spaces that are linear transforms
of R′ G′ B′ . These are those used for the video encoding of television
systems including YIQ and YCC.
In a similar fashion, oRGB first transforms the RGB cube into a parallelepiped via a linear transform. This color space, L′C′C′ , has 3
axes: white-black, yellow-blue, and magenta-red/cyan-green. The
matrix for this transformation is:
 
 ′ 
L
0.2990
0.5870
0.1140 R′
C′  = 0.5000
0.5000 −1.0000 G′ 
(3)
1
′
0.8660
−0.8660
0.0000 B′
C2
and its inverse:
  
R′
1.0000
G′  = 1.0000
1.0000
B′
3.2
0.1140
0.1140
−0.8860
Figure 7: Orthographic view of RGB cube looking along whiteblack axis (left). No affine transform will put red-white-green in
line (middle). The y-axis will have a red-magenta and a green-cyan
as its extreme values (right).
To create a true opponent space where red is actually opposite green
and orthogonal to the yellow-blue axis, we must apply some nonaffine transform. We use the simplest one we can think of: a constant scaling of angle between the opponent directions (see Figure 8). A smoother mapping was attempted, but it made no visual
difference, and so we chose the simplest method that worked well.
R'
# # #
# # #
# # #
(R′ , G′ , B′ )
As can be seen there is no affine transform that will make red,
white, and green collinear. The closest we can come is to put a
red-magenta opposite a green-cyan.
 ′ 
L
0.7436
−0.4111 C1′ 
0.1663 C2′
(4)
oRGB Step 2: Non-uniform rotation around luma
axis L′ from L′C′C′ to oRGB
Is there an affine transform of the RGB cube that moves blue opposite yellow and red opposite green? Unfortunately, there is not.
Figure 7 shows the RGB cube viewed along the white-black axis.
compress
stretch
compress
stretch
Figure 8: To take the red and green directions to the vertical axis, we
compress angles on the blue side, and expand them on the yellow
one.
′ ,C′ )
The transformation from L′C′C′ (L′ ,C1′ ,C2′ ) to oRGB (L′ ,Cyb
rg
is just a compression or decompression of angles depending on
which quadrant the linearly transformed point ends up in. For
points above the yellow-blue axis, the angle in the linear chroma
plane is θ = atan2(C2′ ,C1′ ). The new angle in oRGB space, θo is:
(
(3/2)θ
if θ < π /3
θo (θ ) =
(5)
π /2 + (3/4) (θ − π /3) if π ≥ θ ≥ π /3
′ ,C′ ) in oRGB we simply rotate the
To compute the point (Cyb
rg
′
′
(C1 ,C2 ) point:
′
′ Cyb
C
= R(θo − θ ) 1′
(6)
′
C2
Crg
The inverse transformation angle θ from oRGB to L′C′C′ is:
(
(2/3)θo
if θo < π /2
θ (θo ) =
π /3 + (4/3) (θo − π /2) if π ≥ θo ≥ π /2
(7)
These transformations are symmetric for points below the yellowblue axis. The effect of this rotation along the luma axis on the
gamut of the color space can be seen in Figure 9.
3.3
Properties of oRGB Space
oRGB is a simple R′ G′ B′ -based color space that is useful for computational applications, such as color adjustment, color transformation, and color transfer. It retains the notions of hue and saturation
C'2
C'rg
-1.0
and bottom are cool (see Figure 10). Colors towards the center are
temperate. We show how this simple categorization can be applied
towards cool-to-warm shading models in Section 4.
1.0
0.87
1.0
-1.0
C'1
1.0
C'yb
-0.87
-1.0
Figure 9: Visualization of the gamut of our color space seen along
the luma axis before (left) and after (right) nonuniform rotation.
After rotation, the y-axis is a true red-green axis while keeping the
yellow-blue the same.
presented in HSV, but adds a non-linear perceptual brightness (similar to HSB). This similarity to HSV/HSB makes oRGB a useful
space for color picking. Unlike HSV/HSB, it encodes perceptual
color opponency and its axes encode common color naming used
by people [Hardin 1998]. It also provides a natural computational
framework for categorizing a color by its color temperature - i.e.
warm vs. cool. Finally, the simplicity of the intermediate space
L′C′C′ allows for straightforward gamut mapping that is critical for
computational applications where the manipulations of colors often
produces out-of-gamut results.
3.3.1 Locations of Colors
′ and C′ range in [−1, 1], while the luma
The chroma channels, Cyb
rg
′
channel, L , ranges in [0, 1]. The [L′ , 0, 0] position on the chroma
plane corresponds to a neutral grey color, that is determined by the
luma L′ . In this constant chroma plane, the hue of a color is its angle
in this plane, while its saturation is its distance from the neutral
point.
The oRGB space places the Hering opponent primaries along the
four horizontal and vertical directions. In addition, it places the opponent secondary colors (orange, magenta, cyan, and chartreuse)
exactly between the opponent axes, thus creating an intuitive space
for color modification based on common color naming (see Figure 10).
s
m
ar
w
s
ol
co
Figure 10: Colors to the right of the magenta-chartreuse axis are
warm, to the left are cool, and in the center are temperate.
Studies on basic color naming agree that people across cultures
often refer to colors as warm (red or yellow) and cool (blue and
green) [Berlin and Kay 1969]. oRGB provides a straightforward
computational framework for color selection in terms of color temperature. Colors to the right and top of the magenta-chartreuse di′ axis) are warm, while color to the left
agonal (at 45◦ from the CRG
3.3.2 The oRGB Gamut
The gamut of R′ G′ B′ is the cube defined by its orthogonal basis
vectors - pure red, green and blue. Linear RGB and non-linear
R′ G′ B′ spaces have identical gamut - [0, 1]3 . When we transform
the R′ G′ B′ gamut into L′C′C′ space, its (L′ ,C1′ ,C2′ ) basis vectors are
simply the columns of the matrix in Equation 3. The transformed
gamut space is thus a parallelepiped. Because we use unequal coefficients for determining luma values, L′C′C′ is a true parallelpiped
and not a rotated cube like HSV. Since black is ~0, its mapping into
L′C′C′ is also ~0, therefore the corner of the gamut cube remains at
~0.
Colors that fall within the volume of the transformed gamut space
map to in-gamut colors when converted back to R′ G′ B′ space. Colors that fall outside the volume require gamut mapping.
A good gamut mapping method should first try to preserve luminance and then attempt to preserve hue [Stone et al. 1988]. In
L′C′C′ , luminance is a separate axis, so luminance modification
and preservation are straightforward. Lines of equal angle also represent the same hue, with distance from the grey point being akin
to saturation. These properties allow for simple gamut mapping in
L′ C ′ C ′ .
While our preferred editing space uses a further rotation around the
luma axis of L′C′C′ , projecting along lines of equal hue (and therefore equal angle) is equivalent to performing the same projection in
oRGB. This makes performing gamut mapping for oRGB a straightforward, two step process.
First, we ensure that our luma values are in the range of [0, 1]. To do
so, we generate a piecewise polynomial mapping that ensures that
the mean luma value does not move, while values to the right of the
mean get compressed to fit within [µ , 1], and values to the left of
the mean fit within [0, µ ]:

iβ
h

µ + (1 − µ ) (L−µ )
if L > µ ∧ Lmax > 1
(Lmax −µ )
′
L (L) =
iβ
h

µ − µ (L−µ )
if L ≤ µ ∧ Lmin < 1
(L −µ )
min
where µ is the average luma, Lmin is the minimum luma value,
and Lmax is the maximum luma value. In practice we have found
β = 2/3 to work fairly well.
Second, we need to determine whether the chroma values are outof-gamut, and if so how to bring them into gamut. We found two
approaches that work well in practice - clamping and scaling.
In clamping, we take any pixel values that are outside the L′C′C′
gamut boundary (i.e. the parallelepiped) and project them to the
boundary. To do so, we trace a ray from hL′pixel , 0, 0i to the pixel
value in question, hL′pixel C′pixel C′pixel i. The ray is traveling along
the chroma plane of equal luma for that pixel. If we hit the L′C′C′
parallelepiped before we reach the value, we know that the pixel
value is outside the gamut space and needs to be clamped to the
chroma value found on the intersected boundary. Clamping to the
boundary is already a great improvement over first transforming
from L′C′C′ to R′ G′ B′ and then performing clamping before display
(see Figure 11). This is because clamping to the boundary in L′C′C′
preserves both the luma and the hue of a particular value, while
clamping an R′ G′ B′ triple to [0, 1] does not.
(a) Original
(b) Clamp in R′ G′ B′
(c) Clamp in L′C′C′
(d) Scale in L′C′C′
Figure 11: The original image is modified by doubling the standard deviation in the red-green channel resulting in out-of-gamut
red values but improved green variation. Clamping in R′ G′ B′ leaves
overly saturated reds. Both clamping and scaling in L′C′C′ preserve
luma, while scaling also produces a more even saturation distribution. Both methods retain the improved green variation.
green
red
as well as for selection of colors for shading models.
blue
yellow
If many pixel values are outside of the L′C′C′ gamut boundary,
however, clamping in L′C′C′ will produce a very saturated image, as many values will be truncated to the most saturated part
of the space. For this case, we scale values towards the center, i.e.
hL′ , 0, 0i, instead of simply clamping to the boundary. To do so, we
first divide the parallelogram into planes of equal luma. For each
equal luma plane, we discretize the plane into many angular slices.
For each angular slice, we compute both the furthest distance from
the grey point and the closest point on the boundary for all pixel
values that fall within the slice. We then project all pixel values
within that slice so that the point furthest out will lie on the boundary of the parallelepiped. To avoid color discontinuities we use
about 3000 angular slices per luma plane which result in visually
smooth mappings of the colors.
Figure 12: Modifying color contrast through deviation adjustment.
One step down/left scales deviation by 1/α while a step up/right
scales deviation in that channel by α . The horizontal axis is a
yellow-blue standard deviation change (left lowers and right increases) and the vertical axis is a red-green change (down lowers
and up increases).
Using a simple linear scaling provides excellent results in practice
(see Figure 11), though it is possible that some more complicated
non-linear method may be necessary in some cases. Gamut mapping via scaling is used for all oRGB results in the paper, unless
otherwise noted.
Sometimes, it is more useful to simply take an existing image and
make it “warmer” or “cooler” in tone. This operation is straightforward in oRGB, as the two chroma channels provide a direct mapping into a color’s temperature. Since the two chroma channels are
independent, we can adjust them without introducing global image
color casts (see Figure 14).
4
Sample Applications
We now demonstrate the utility of the oRGB color space in four different applications: color adjustment, cool-to-warm shading, color
transfer, and gamut mapping. For color adjustment, we show how
oRGB’s natural opponent axes can be used to provide a meaningful Photoshop-style “variations” tool. As oRGB’s natural primaries
and secondaries are intuitive they can aid in color selection and
can serve as a natural metric for color temperature in cool-to-warm
shading. For color transfer, we demonstrate that oRGB performs at
least as well as previously suggested computational spaces. Finally,
we examine gamut mapping in the context of one computational application - color transfer.
4.1
Color Adjustment
Users want to perform simple, yet intuitive color adjustments. Here
we show how oRGB can be used for Photoshop-style “variations”
Our color space is useful for simple color contrast adjustments as
we can increase the saturation of the red-green channel independently from the yellow-blue one. To do so, we simply have to scale
the standard deviation in each channel separately. This operation
provides useful and intuitive variations results (see Figures 12 and
13).
We can also apply the concept of “variations” to material parameters. Varying the diffuse texture via mean shifting in the chroma
plane generates meaningful alternative materials (see Figure 15).
Shifting up produces a redder tone, up and right - an orange tone,
right - a more yellow tone, and right and down - a chartreuse tone.
Similarly, shifting left and up produces a magenta tone, left - a bluer
tone, left and down - a cyan tone, and down - a greener tone.
4.2
Cool to Warm Shading
Gooch et al. [Gooch et al. 1998] present a cool-to-warm shading
model that maps surface normal to color, with normals pointing
away from the light being associated with cool colors, and those that
point toward the light being warm. They also present a model for
choosing the colors based on intrinsic surface color. Their model
is:
C(cos) = (1 − f (cos))Ccool + f (cos)Cwarm ,
(8)
where cos is the cosine between surface normal and a light direction, f is a monotonically increasing function that maps [−1, 1] to
red
green
red
green
blue
blue
yellow
Figure 13: Another example of color adjustment using deviation
scaling, similar to Figure 12.
[0, 1], and Ccool and Cwarm are empirically chosen colors. They
present results for a linear f and a variety of Ccool and Cwarm .
A difficulty that arises when using this model is that there is no
metric that helps users decide how “warm” or “cool” a color is. This
issue can be directly addressed in oRGB (see Figure 10), as a color
is warm if its dominant direction is toward the red or yellow axes,
and cool if its dominant direction is toward the blue or green axes.
′ ,C′ ) · (0, 1, 1) > 0
Quantitatively a warm color is one where (L′ ,Cyb
rg
and a cool color is one where that dot product is negative.
To determine whether the intuitive axes of the oRGB space are helpful in NPR applications, we first added a cool-warm test to an NPR
shader that uses Equation 8. We also explored changes to the interpolation rule in an attempt to make NPR shaded objects less flat
looking. If the three channels are interpolated separately, then we
have:

 
 ′ 
f0 (cos)Lw
L
(1 − f0 (cos))Lc
w
c
′ 
Cyb
(9)
= (1 − f1 (cos))Cyb  +  f1 (cos)Cyb 
c
w
′
(1 − f2 (cos))Crg
f2 (cos)Crg
Crg
c ,Cc ) and (Lw ,Cw ,Cw ) are the oRGB components
where (Lc ,Cyb
rg
yb rg
of Ccool and Cwarm respectively. This will be a valid cool to warm
mapping provided the end colors are warm and cool, and that all
of fi are strictly non-decreasing and fi (−1) = 0, fi (1) = 1). If the
two chroma channels are handled as in Gooch et al. [Gooch et al.
1998] and the luma channel is an average between linear interpolation and cosine-based interpolation (Gouraud lighting), then an
NPR look can be retained while also having some visual character
of diffuse lighting (Figure 16). Our point is not that this is a better
yellow
Figure 14: Changing the overall tone of an image through mean
shifting. The natural axes of oRGB provide intuitive shifts along
primaries (yellow, red, blue, and green) and secondaries (orange,
magenta, cyan, and chartreuse). The horizontal axis is yellow (shifting right) and blue (shifting left). The vertical axis is red (shifting
up) and green (shifting down).
NPR model, but that oRGB provides a space where reasoning about
NPR model design is fairly natural.
4.3
Color Transfer
For color transfer, Reinhard et al. [Reinhard et al. 2001] demonstrated that the l αβ color space [Ruderman 1998] provides excellent results when compared to the RGB color space. While l αβ
works well in many situations, oRGB provides features we feel are
lacking in l αβ . For example, color transfer in oRGB allows for
independent manipulation of the luma, pure yellow-blue, and pure
red-green channels. Thus, the statistical properties of the channels
can be transfered in a space that is meaningful in terms of natural
primaries and secondaries.
We compare color transfer results (see Figures 17 and 18) produced by oRGB with those of three different Hering-style opponent
color models - Schwarz [Schwarz et al. 1987], RBW [Shirriff 1993],
and RGBY [Ware and Cowan 1990]. We show oRGB results that
use no gamut mapping or use gamut scaling. In addition, we compare against Adobe Photoshop’s “Color Match” function, as well as
HSV, l αβ , CIE L*a*b* and YCC.
In oRGB shifting red towards blue produces purple, while retaining
yellow (Figure 17), and shifting magenta towards red results in true
red (Figure 18). The opponent spaces we compare against also have
orthogonal axes, and the hue shifts in these spaces seem to produce
red
4.4
Gamut Mapping
green
Gamut mapping is a necessary operation for computational applications, as values are often shifted and scaled, thus increasing the
chance for producing out-of-gamut pixel values. Because oRGB is
a linear transform of R’G’B’ followed by a rotational distortion, the
gamut boundary is simpler than for most other spaces. While this
does not necessarily guarantee fewer out-of-gamut values, it allows
us to correct such cases more easily than in CIE L*a*b* or l αβ .
blue
yellow
Figure 15: Instead of applying mean shifting to the final image, as
in Figure 14, we instead used the intuitive oRGB axes as a color
“variations” selector for the diffuse term in a car rendering.
One computational application, color transfer, works by shifting
and scaling groups of pixels without regard to gamut boundaries.
While any image used for transferring statistics will have a well
defined mean in the RGB gamut space, some values towards the
tails of the distributions may land outside the RGB gamut following
transfer.
To provide a comparison metric for the out-of-gamut results produced by different color spaces during color transfer, we did the
following test. We took the top 12 images produced by Google images, using “family vacation” as a search term. We then did color
transfers of the 12 photos to each other, producing 132 non-trivial
transfer results per color space. Table 1 summarizes the averaged
results for each of the color spaces.
Color Space
oRGB
HSV
l αβ
Schwarz
RBW
RGBY
CIE L*a*b*
YCC
Figure 16: Bottom: shading from Gooch et al. Top: the luma channel in oRGB is modified to partially resemble traditional Gouraud
style shading while maintaining a cool-to-warm transition.
reasonable results. However, the saturation and brightness in these
spaces do not seem to perform as well, perhaps as a result of the
fact that they operate in linear space. In Figure 19 we show the
importance of running in non-linear space for oRGB. The rest of the
spaces do not have orthogonal red-green and yellow-blue axes, and
therefore independent shifting and scaling results in mixing some
of the color from the other axis (e.g. the magenta cast for l αβ in
Figure 18).
% Pixels
10.36
9.07
10.46
16.45
11.95
13.67
10.89
11.24
%R
22.58
121.87
104.54
>999.9
146.29
296.99
117.82
22.99
%G
16.59
122.12
40.56
>999.9
124.20
234.07
66.05
15.99
%B
26.42
118.9
25.02
>999.9
154.94
174.70
38.33
25.91
Table 1: Different color spaces tend to produce different out-ofgamut results. The first column lists the color space tested. The
second column shows the average percentage of out-of-gamut pixels. The remaining three columns demonstrate the out-of-gamut
percentage range per RGB channel.
CIE L*a*b* and l αβ produce a similar number of out-of-gamut
values to oRGB, however, they tend to generate results that are further from the RGB gamut boundary. For these spaces, simply clipping once the color transfer result is returned to RGB can be disastrous. While oRGB has very low out-of-gamut values by comparison, we still seek to remove these remaining out-of-gamut values
via gamut mapping.
Figure 19: Linear version of our model for color transfer (original
and target images are the same as the ones used in 17 and 18).
It should be noted that orthogonality is not the full story for color
transfer. RGB has orthogonal axes, but unfortunately for natural
images, it has highly correlated axes restricting its use for color
transfer [Reinhard et al. 2001]. While l αβ produces de-correlated
axes using PCA, oRGB attempts to utilize opponent-color perceptual axes for meaningful computational applications.
Color transfer is a complex operation that is highly dependent on
the specific image distributions. We don’t believe there is one color
space that can consistently provide the best results. However, we
believe oRGB performs at least as good, and in many cases - better,
than other computational color spaces, based on the set of images
we’ve tested.
Having the ability to perform gamut mapping should not be ignored. The l αβ and CIE L*a*b* results in Figure 17 produce results that are wildly out of the RGB gamut, thus resulting in unpredictable and unexpected color transfer results.
5
Conclusion
The oRGB color space we have introduced has two basic merits.
First, it is a relatively simple transformation of R’G’B’, so it is efficient, easy to implement, and has a reasonably simple gamut. Second, it has orthogonal axes that are aligned with Hering’s opponent
color pairs.
We have shown empirically that this space can be useful in some
common computer graphics computations. We suspect the space
works well for three reasons. First, the gamma encoding of
(R′ , G′ , B′ ) values serves the same role as the non-linearities in models such as CIE L*a*b*. Second, the simple gamut boundary reduces out-of-gamut issues and allows any that remain to be easily
corrected. Third, we believe that for computer graphics applications, there is something natural about the Hering primaries and
their opponency.
There are several limitations of the oRGB space. As a computer
graphics space, it is not well-suited to traditional colorimetric applications where all possible colors, as opposed to all displayable
colors, are represented. In applications where the Hering primary
colors are not central, it is unlikely to have advantages over linear
transformations of R’G’B’ such as YCC. Perhaps its biggest limitation is that it is not a pure bijection of RGB, so it is not the perfect
space for choosing colors for RGB displays.
S HIRRIFF , K. W. 1993. The RBW color model. Computers and
Graphics 17, 5, 597–602.
S MITH , A. R. 1978. Color gamut transform pairs. In Proceedings
of SIGGRAPH, 12–19.
S TONE , M. C., C OWAN , W. B., AND B EATTY, J. C. 1988. Color
gamut mapping and the printing of digital color images. ACM
Trans. Graph. 7, 4, 249–292.
VALBERG , A. 2005. Light, vision, color. Wiley.
VOLBRECHT, V. J., AND K LIEGL , R. 1998. The perception of
blackness: An historical and contemporary review. In Color Vision: Perspectives from Different Discplines, W. G. K. Backhaus,
R. Kliegl, and J. S. Werner, Eds. De Gruyter, 187–206.
WARE , C., AND C OWAN , W. 1990. The rgyb color geometry.
ACM Trans. Graph. 9, 2, 226–232.
References
B ERLIN , B., AND K AY, P. 1969. Basic Color Terms. Their Universality and Evolution. Berkeley: University of California Press.
Reprinted 1991.
C HICHILISKY, E., AND WANDELL , B. A. 1999. Trichromatic
opponent color classification. Vision Research 39, 20, 3444–58.
C ONWAY, B. 2002. Neural Mechanisms of Color Vision. Kluwer
Academic Publishers.
FAIRCHILD , M. D.
Addison-Wesley.
2005.
Color Appearance Models, 2 ed.
G OOCH , A., G OOCH , B., S HIRLEY, P., AND C OHEN , E. 1998. A
non-photorealistic lighting model for automatic technical illustration. In Proceedings of SIGGRAPH, 447–452.
H ARDIN , C. L. 1998. Basic color terms and basic color categories. In Color Vision: Perspectives from Different Discplines,
W. G. K. Backhaus, R. Kliegl, and J. S. Werner, Eds. De Gruyter,
207–218.
H UNT, R. 2004. The Reproduction of Colour. Wiley.
H URVICH , L., AND JAMESON , D. 1957. An opponent-process
theory of color vision. Psychol Rev 64, 384–404.
K UEHNI , R. G. 2003. Color Space and its Divisions. Wiley.
K UEHNI , R. G. 2004. Color: an introduction to practice and
principles, second ed. Wiley.
M OROVIC , J., AND L UO , M. R. 2001. The fundamentals of gamut
mapping: A survey. Journal of Imaging Science and Technology
45, 3, 283–290.
NAIMAN , A. C. 1985. Color spaces and color contrast. The Visual
Computer 1, 3, 194–201.
P OYNTON , C., 2004. Color science and color appearance models
for cg, hdtv, and d-cinema. SIGGRAPH course notes.
R EINHARD , E., A SHIKHMIN , M., G OOCH , B., AND S HIRLEY, P.
2001. Color transfer between images. Computer Graphics &
Applications, 2–8.
RUDERMAN , D. L. 1998. Statistics of cone responses to natural
images: Implications for visual coding. J. Opt. Soc. Am. A 15, 8
(August), 2036–2045.
S CHWARZ , M. W., C OWAN , W. B., AND B EATTY, J. C. 1987.
An experimental comparison of rgb, yiq, lab, hsv, and opponent
color models. ACM Trans. Graph. 6, 2, 123–158.
(a) Original Image
(b) oRGB none
(c) oRGB scale
(d) HSV
(e) Photoshop
(f) l αβ
(g) Target Image
(h) Schwarz
(i) RBW
(j) RGBY
(k) CIE L*a*b*
(l) YCC
Figure 17: The statistics of the target image are transferred to the original image in a variety of color spaces. Transferring red towards blue
produces purple in oRGB instead of yellows, reds, or browns produced by the other spaces.
(a) Original Image
(b) oRGB none
(c) oRGB scale
(d) HSV
(e) Photoshop
(f) l αβ
(g) Target Image
(h) Schwarz
(i) RBW
(j) RGBY
(k) CIE L*a*b*
(l) YCC
Figure 18: The statistics of the target image are transferred to the original image in a variety of color spaces. A magenta shifted towards pure
red becomes pure red in oRGB instead of remaining magenta.